[bugfix] matrixgame2 causal: build block-mask cache with effective block size

The cached BlockMask was keyed/built from self.num_frame_per_block, but the
transformer blocks run with effective_num_frame_per_block (the per-call
num_frame_per_block override). Align the mask with the block size the blocks
actually use; it is part of the cache key so different sizes get distinct
entries. Verified on H100: the size-1 and size-3 masks genuinely differ, so
any forward whose effective size diverges from the default (e.g. a boundary
block of size 1) would otherwise apply the wrong attention pattern. For the
current shipped matrixgame2 configs the two sizes coincide, so output is
unchanged — this removes the latent mismatch.

Author: rich7420

fastvideo/models/dits/matrixgame2/causal_model.py

@@ -1029,7 +1029,9 @@ def _forward_inference(
             self._block_mask_cache = bm_cache
         _dev = hidden_states.device
         _fsl = post_patch_height * post_patch_width
-        _nfb = self.num_frame_per_block
+        # Use the block size the blocks actually run with (the per-call override),
+        # not the model default, so the mask matches. _nfb is in the cache key.
+        _nfb = effective_num_frame_per_block
         _las = self.local_attn_size
 
         _k = ("main", num_frames, _fsl, _nfb, _las, _dev)